[CRIU] [PATCH] restorer: block signals before restoring SIGCHLD handler

[Pavel Emelyanov]

On 09/20/2013 07:09 PM, Andrew Vagin wrote:
> On Fri, Sep 20, 2013 at 06:59:33PM +0400, Pavel Emelyanov wrote:
>> On 09/20/2013 06:21 PM, Andrey Vagin wrote:
>>> Otherwise the process can handler this signal before calling sigreturn.
>>>
>>> And threads must block signals before switching on
>>> CR_STATE_RESTORE_SIGCHLD, because signals are restoring on this stage.
>>>
>>> So if a thread leader starts restoring shared signals before another
>>> thread has blocked signals, this thread can start to hadle the signal.
> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> The previous two lines describe the race
> 
> __export_restore_task				| __export_restore_thread
> 
> restore_finish_stage(CR_STATE_RESTORE);		| restore_finish_stage(CR_STATE_RESTORE);
> sys_sigaction(SIGCHLD, &args->sigchld_act, ...) |
> restore_signals()				|
> 	ksigfillset(&to_block);			|
> 	sys_rt_sigqueueinfo(sys_getpid(), ...); |
> 						| execute signal handler() <------ ALARM !!!
> 						| restore_signals()
> 						|	ksigfillset(&to_block);
> 
>>> I have seen this race in static/sigpending.
>>
>> What race? Plz, elaborate.
> 
>>
>>> Signed-off-by: Andrey Vagin <avagin at openvz.org>

ok, applied